Meanwhile, DeepSeek has slashed prices on one of its models by 75% and it’s starting to get some traction. Never mind that your data literally gets hosted on a PRC server… What could go wrong?
In other words, they want all of their competition to stop competing with them for a while.
“At least until we’re done working on this IPO!”
In another “meanwhile”:
The report said some employees at [Chinese] firms including Alibaba and DeepSeek are now required to obtain approval from authorities before travelling abroad.
It’s the Manhattan Project all over again!
Do you really think this is why Anthropic said it? They’re one of the leaders and a pause could only hurt their lead.
Or do you think it was for marketing before the IPO?
Anthropic has been the most outspoken, for-profit on AI risk. If I have my facts right, Dario Amodei, was one of the researchers that flagged the danger back with GPT-2 or GPT-3.
And recently this:
Also in February 2026, Anthropic donated $20 million to the nonprofit organization Public First Action, which has announced plans to support political candidates that support regulation of AI.[30]
Cynically you can say they want to influence AI regulation by funding it, but I’m not convinced it isn’t genuine.
When a company whose sole purpose for existence is AI research says that they want a pause on AI research, that’s really the only interpretation that’s possible. If they thought a pause in the research was actually needed, they wouldn’t be in the business they’re in at all. They want everyone else to pause, while they keep on researching.
I don’t think so. As I said, a pause hurts industry leaders because it lets competition catch up.
Possible cynical interpretations are:
- They want to brand themselves as the responsible AI company
- They want to influence AI regulation to create a higher barrier of entry against competition.
- They don’t have the compute to scale. They’re currently leasing capacity from xAI and Google.
On the flip side, they held off releasing their Mythos model and shared the preview version with financial institutions due to cyber security concerns. That doesn’t help their business much unless it’s an elaborate branding campaign.
I would add that the pause has no chance of happening and they know that. Likely the motivation is sincere or subtley cynical.
Only if the industry leaders actually do it. If they intended to pause, they’d already be doing it. A “call for a pause” just means they want everyone else to.
I had thought DeepSeek’s supposed non-reliance on NVIDIA tech being too pricey or unavailable could be a game changer. Yet NVIDIA, still important in BTC mining, went seamlessly to AI.
I use ChatGPT in development rather extensively. I’d have backed off from Ubuntu 26.04 unless ChatGPT could assist in installing stuff like Mongo DB. “We’ve” agreed to wait till the probable .1 Ubuntu release in a month or so to get even my local system working well, let alone deploy it remotely.
Other than that, I like OpenRouter for allowing some basicaly ‘free’ models to be available with up to 1000 requests.
I admit, though, I haven’t had to make decisions on professional level AI/LLM’s.
I have now read the Anthropic think piece, and I’d say the media is misdescribing the content, and that 1. and 2. are correct. The paper says a pause would be nice, and it would have to be mediated by big companies (I wonder who?), but it requires further research before even being actually proposed.
It is also a massive puff piece about its own coding awesomeness.
Everything has to be viewed as IPO table-setting at this point.
With a wave of IPOs potentially coming, the market is likely going to take note of the IPO for Cerebras last month. Cerebras builds big processors for AI, and came into the market at $185…at which point the value settled up by an additional 68% (note: whenever you have your own personal IPO, find out who their advisor was and don’t hire them).
It feels like a ton of capital is just poised, waiting to invest in that first shiny object. And if Anthropic is first, a $trillion+ valuation seems pretty evident.
And apparently OpenAI has filed as well.
I ran across that news from this article on a potential price war among AI copmanies:
There isn’t much in that link since it is an article about a WSJ article, but here is the gist:
Meanwhile, as the Wall Street Journal reports, executives at OpenAI are pondering whether to kick off a price war with the company’s biggest competitor, Anthropic. By dramatically lowering prices, the company’s reportedly hoping to steal users, while also anticipating similar price cuts by its competitor.
Cloudflare announced a beta program for a new feature for their AI Gateway (AI proxy) service last week. The goal is to provide better spending control for AI usage. It shores up functionality that is missing from some AI providers to limit usage to budgeted limits. You can set budgets at different levels of granularity (by team, by person, etc.).
What I think is really interesting is that for the next phase, they want to add logic to detect the task and route to an appropriate model; summarizing emails would go to a smaller, cheaper model than writing software.
The biggest impact though is that a company using an AI Gateway is not locked into a single AI provider. A gateway can surf the AI price wars and route to the cheapest models.
Here is their announcement:
Cloudflare isn’t the only company providing AI gateway services or products. There are several free or open-source projects that are robust. Of note, one popular Python-based project, LiteLLM, suffered an exploit in March. One of the official releases had an exploit that leaked the API keys to the hackers:
That was really interesting, thanks for sharing it. It makes sense as a business model; I expect it would be largely invisible to the LLM providers (I think since the ‘router’ would be calling the APIs), but I wonder how they’d feel about it.
There are a few products on the market for AI gateways and Google bought a startup that has a nice solution. You can proxy your requests through the gateway and it has an API adaptation layer to convert version your client’s and serve’s APIs. This makes it pretty easy to switch AI providers (or do load balancing, billing, etc.)
I was thinking that too. I think the way for them to recover some vendor leverage is to create better tools/agents so they run directly on client’s workstations or in web consoles. We see this too with coding agents, copilot, and Google docs integration.